1 Crowdedness and Mobile Targeting Effectiveness Michelle ANDREWS, Xueming LUO, Zheng FANG, and Anindya GHOSE March 2014 Advances in mobile technologies can provide novel measures of the crowdedness of a consumer’s immediate social environment. Our paper seeks to understand how crowdedness affects consumer response to mobile targeting. This unconventional but important question is broadly related to how engagement with mobile content changes in contextual environments. We rely on a unique marketing environment –subway trains– in which the crowdedness of the environment is observable to empirically examine this question. With the cooperation of a leading mobile telecom provider, we provide targeted messages to subway passengers. On the bases of multi-method field data, we find robust evidence that response behavior to mobile targeting varies as a function of crowdedness in the trains. The evidence relies on exploiting an exogenous shock due to a traffic intervention, establishing user homogeneity via same-train-same-time commuters and propensity score matching, using a residual crowding approach and multiple falsification tests, and conducting additional checks with field surveys. According to the field surveys, consumers’ immersion in their mobile phones explain why crowdedness in subway trains affects mobile involvement and ultimate purchases. Marketers may consider gauging the crowdedness of a consumer’s social environment as a new way to boost mobile targeting effectiveness. Key words: mobile targeting, crowdedness, field study, multi-method, new technology
33
Embed
Crowdedness and Mobile Targeting Effectiveness · question with broad implications for mobile targeting in the context of consumers’ social environment. Second, methodologically,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Crowdedness and Mobile Targeting Effectiveness
Michelle ANDREWS, Xueming LUO, Zheng FANG, and Anindya GHOSE
March 2014
Advances in mobile technologies can provide novel measures of the crowdedness of a consumer’s
immediate social environment. Our paper seeks to understand how crowdedness affects consumer
response to mobile targeting. This unconventional but important question is broadly related to how
engagement with mobile content changes in contextual environments. We rely on a unique marketing
environment –subway trains– in which the crowdedness of the environment is observable to empirically
examine this question. With the cooperation of a leading mobile telecom provider, we provide targeted
messages to subway passengers. On the bases of multi-method field data, we find robust evidence that
response behavior to mobile targeting varies as a function of crowdedness in the trains. The evidence
relies on exploiting an exogenous shock due to a traffic intervention, establishing user homogeneity via
same-train-same-time commuters and propensity score matching, using a residual crowding approach and
multiple falsification tests, and conducting additional checks with field surveys. According to the field
surveys, consumers’ immersion in their mobile phones explain why crowdedness in subway trains affects
mobile involvement and ultimate purchases. Marketers may consider gauging the crowdedness of a
consumer’s social environment as a new way to boost mobile targeting effectiveness.
Key words: mobile targeting, crowdedness, field study, multi-method, new technology
2
1. Introduction
Mobile marketers have a vested interest in learning how to optimize their targeting campaigns. As mobile
ad spending is projected to exceed $60 billion by 2017 (eMarketer 2013), marketers are keen on
understanding what impacts consumer responses to mobile promotions. The effectiveness of mobile
targeting relies on reaching consumers when and where they are most receptive to marketing messages.
Recent examples include geo-fencing (i.e., sending mobile coupons to people within the virtual perimeter
of a store) and iBeacon (i.e., transmitting geo-located deals to within-store devices). Several studies have
investigated the contextual factors that influence the effectiveness of mobile targeting strategies. Mobile
users’ internet search behavior (Ghose et al. 2013), cross platform synergies with web advertising (Ghose
et al. 2014), geographic mobility (Ghose and Han 2011), location and time at which they receive a
promotion (Luo et al. 2014), distance between the store and the user (Molitor et al 2014), the weather
(Molitor et al. 2013), and the product characteristics (Bart et al. 2014) play an important role in driving
mobile purchase likelihood.
This work seeks to understand how crowdedness affects response to mobile targeting in the context of
consumers’ social environment.1 We use a unique marketing environment –subway trains– in which the
crowdedness of the environment is directly observable via mobile technology. The subway environment
is a promising and relevant marketing environment because in most cities people spend a considerable
amount of time commuting, averaging 48 minutes each way, according to Census Bureau reports. More
specifically, the underground subway in the city of our study is mobile-equipped and allows passengers to
use their mobiles throughout their commute. With the cooperation of a leading mobile telecom provider,
we provided targeted short message services (SMSs) to consumers in subway trains. Recipients could
purchase the promoted service by responding to the SMS. The natural setting of subway trains with
different levels of crowdedness enables us to examine the effects of crowdedness. We use cellular
technology to record in real-time the number of mobile users located within each mobile-equipped
subway train. Our knowledge of the exact dimensions of the subway train enables us to determine
people’s spatial proximity to one another. We thus gauge crowdedness as the number of subway
passengers per square meter and test how crowdedness impacts mobile targeting effectiveness.
Our identification strategy relies on multi-faceted field data. Since it is extremely difficult to
manipulate crowdedness in a real world setting, multi-method approaches are critical to reduce the threats
of self-selection, endogeneity, and other potential confounds. Commuters may self-select into more or
less crowded trains as a function of their work schedule demands. The ideal test of the effects of
crowdedness would be to randomly assign commuters to crowded versus non-crowded trains, but this is
practically impossible to accomplish. Therefore, to reduce sample-selection concerns and establish
1 We define crowdedness as the physical social presence of others. We use crowdedness and crowding interchangeably.
3
commuter homogeneity, we rely on five key strategies. First, our field data involves both peak and non-
peak hours with SMSs sent to different trains from the morning to the evening. Thus, we capture various
time cycles and control for systematic differences between different types of subway commuters. Second,
our field data comes from two different days: a weekday and the weekend. This is useful because
travelers during the weekdays may differ from those during the weekend, and this helps reduce concerns
of selection bias. Third, our corporate partner randomly sampled commuters from amongst our targeted
subway population using an instant computational and randomization procedure. This randomization
helps alleviate concerns about mobile users’ heterogeneity. Fourth, we demonstrate that in our field data
individuals are virtually similar in all aspects of mobile usage behavior. We do so by controlling for the
potential confounds of each individual’s mobile usage in terms of their monthly average revenue, call
time, messages, and data usage. In addition, we use same-train-same-time commuters and propensity
score matching methods to assure that we can isolate the impact of crowdedness conditional on commuter
homogeneity. Fifth, a further potential confound would be the existence of an unobserved variable that
drives both crowdedness and response to mobile targeting. To reduce this possible endogeneity threat, our
field study exploits an exogenous increase in crowdedness as a result of an unexpected traffic intervention
enforced by the government. This intervention was brought about by a high-security police escort for an
important politician, which created an exogenous spike in crowdedness within the subway trains. These
five steps help reduce sample selection threats and boost the confidence of the internal and external
validity of the results.
Therefore, these measures enable us to isolate the effect of crowdedness on mobile purchase behavior
when comparing mobile users in high versus low crowded environments. We find that consumer
responses vary by the level of crowdedness in the trains. In congested subway trains, purchase rates were
significantly higher than in uncongested ones. That is, consumers in more (versus less) crowded trains
have a higher likelihood of responding to targeted mobile promotions, conditional on commuter
homogeneity.
To furnish more robust evidence, we augment the field data with field surveys. Through its customer
call center, our corporate partner contacted mobile users who received and purchased the targeted
promotion, as well as those who received but did not purchase it. We matched the attitudinal surveys with
observed crowdedness and mobile purchase records. We not only confirm the effects of crowdedness on
purchases, but also explain that such effects arise at least partly due to mobile immersion. That is, in
crowded subway trains, commuters experience a loss of physical space. To psychologically cope with this
spatial loss and avoid accidental gazes, commuters escape into their personal mobile space (i.e., mobile
immersion). In turn, via this immersion, passengers become more involved in targeted mobile messages,
and consequently more likely to make a purchase. Also, the surveys help rule out alternative explanations
4
(such as social anxiety, the desire to show off the latest smartphone devices, price sensitivity, and deal
proneness). Thus, we add further corroborating evidence with attitudinal measures that are unobserved in
the field data but observable in the field surveys.
Our contributions to the literature are two-fold. First, substantively, we investigate the unconventional
question of whether social crowding changes consumer behavior in the context of mobile targeting. To
our knowledge, no one has explored the effects of crowdedness on mobile purchases in the unique
marketing environment of subway trains. Our finding that crowdedness can positively affect mobile
purchase behavior is counter-intuitive. Indeed, much of the literature on crowdedness has largely
demonstrated negative effects, including avoidance behaviors and risk-aversion (Harrell et al. 1980;
Maeng et al. 2013). In the context of our subway setting, we find that targeting consumers in crowded
trains with mobile promotions may serve as a welcome relief. Hence, we enhance our understanding of
how crowdedness affects consumer receptivity to mobile promotions, which is an uncommon but pivotal
question with broad implications for mobile targeting in the context of consumers’ social environment.
Second, methodologically, we rely on a rigorous multi-method research design with field data and
attitudinal surveys. Traditionally, crowdedness is often difficult and impossible to measure in
conventional retail environments. Today, mobile technologies allow researchers to more precisely
measure crowdedness in real-time. Our field data is interesting due to the unique environment where
consumers can be individually targeted for promotions, and the level of crowdedness that can be
measured with a great deal of precision with modern mobile technology. We identity the effects of
crowdedness by controlling for self-selection across peak versus non-peak, weekday versus weekend, and
exogenous versus non-exogenous crowdedness. Also, our field surveys shed light on additional evidence
for why consumer response to mobile targeting varies as a function of crowdedness, evidence that is
based on consumer attitudes from the surveys and not directly observed in the field data.
Our findings have meaningful implications for mobile targeting practices. For managers, crowding
represents a novel consumer setting for gauging the effectiveness of context-sensitive mobile messages.
Because consumers may be more engaged with their devices in a crowd, such social environments may be
a good thing for marketers who can deliver information at the right moment. More specifically, public
transit commutes provide a unique marketing environment and thus a window of opportunity for
marketers to target mobile users. As cities facilitate mobile usage in underground subways (Flegenheimer
2013), and since consumers spend substantial amounts of time commuting during quotidian life,
marketing opportunities to target customers based on ambient factors such as crowdedness will abound.
5
2. Literature on Mobile Marketing and Crowdedness
Our research builds on two streams of research. First, the literature on mobile marketing across
marketing, information systems, and management science is relatively nascent. Ghose et al. (2013) find
that consumers are more likely to click on links to stores located close to them and on higher-ranked links
on their mobile screens. This is due to consumers’ tendency to browse the mobile internet for their
immediate, local needs and the higher search costs associated with the smaller screen sizes of mobile
devices. Indeed, building on that study, in a study of location-based advertising with smartphone users,
Molitor et al. (2014) found that mobile coupon redemption rates increased the closer consumers were to a
store and higher the offer was displayed on the screen, conditional on the discount. Echoing this, Luo et
al. (2014) show how the location of consumers relative to a promoted venue and the time at which they
receive a promotion affects their mobile purchase likelihood due to the contextual benefits of having
enough planning and travel time. In terms of smartphone users’ geographic mobility, Ghose and Han
(2011) find that people are more likely to use mobile content than to generate it when they are travelling,
and that the variance of user’s traveling patterns is a stronger predictor of their propensity to engage with
the mobile device than the mean. Researchers have demonstrated how real-time mobile promotions
transmitted within grocery stores can increase shoppers’ travel distance and consequently boost their
unplanned spending (Hui et al. 2013). Scholars have also found that for mobile display advertising,
products higher on involvement and utilitarian dimensions generate more favorable attitudes and purchase
intentions (Bart et al. 2014). Another emerging stream of work uses randomized field experiments to
causally measure cross-platform synergies between web and mobile advertising (Ghose et al. 2014).
Researchers demonstrated that in-app advertisements decrease consumer demand for mobile apps (Ghose
and Han 2014). Environmental factors may also explain why consumers make mobile purchases. For
instance, Molitor et al. (2013) show how the weather impacts consumers’ mobile coupon choice.
Following this stream of research, we suggest the effectiveness of mobile targeting is context dependent.
A particular context in which mobile users may find themselves –often several times a day– is that of a
crowd. People may experience crowdedness while commuting, buying food, and shopping.
Second, our work is related to the literature on crowdedness. In sociology and psychology, studies
have linked crowdedness to physical and mental diseases and juvenile delinquency (Schmitt 1966).
Crowdedness can increase stress (Collette and Webb 1976), frustration (Sherrod 1974), and hostility
(Griffitt and Veitch 1971). Individuals may consider themselves more anonymous in crowds, which can
reduce their social interactions and fuel antisocial behavior (Zimbardo 1969). In more crowded areas, for
example, crime rates were found to spike due to the higher likelihood that criminals could avoid detection
(Jarrell and Howsen 1990). Scholars also suggest that in more crowded settings, people perceive less
control over their situations (Aiello et al. 1977). In the consumer behavior research, crowdedness has been
6
found to induce avoidance behaviors and shorter shopping times (Harrell et al. 1980). Other research has
shown that crowding can threaten consumers’ sense of uniqueness, which may lead them to purchase
more distinctive products in order to restore their perceived individuality (Xu et al. 2012). Recently,
Maeng and colleagues (2013) demonstrated that crowded environments can spur shoppers to become
more risk averse and favor safety-oriented options such as visiting a pharmacy over a convenience store.
In our study, we integrate and advance these two streams of research on mobile marketing and
crowding. Specifically, our multi-method field data can precisely quantify crowdedness with mobile
technologies. Holding commuter type constant in terms of homogeneous mobile usage, we empirically
test whether consumer purchase behavior varies as a function of the level of crowdedness.
3. Field Data
To explore how crowdedness affects consumer response to mobile targeting, we use data from a field
study conducted in the subway system of a major city with the cooperation of a leading mobile telecom
provider (who wishes to remain anonymous). Our targeted population consisted of mobile users from a
city with a population of 20 million. Our field data comprises four parts of data, from 1) a business
weekday, 2) a weekend day, 3) exogenous crowdedness, and 4) follow-up surveys.
3.1 Method
In parts 1 and 2 of our field data, a total of 10,690 mobile users who rode the subway were randomly
selected to be sent a promotion through a text message and their responses were recorded. The SMS
advertised a missed call alert service, which if purchased would notify mobile users of the time and
number of the calls they missed. Our study took place in China in September 2013, where cell phone
plans are piecemeal rather than all inclusive. As such, mobile users must purchase a missed call alert
service separately in order to be notified of missed calls. The SMS read “Missed a call and want to know
from whom? Subscribe to [Wireless Service Provider’s] missed call alert service and be notified by SMS
of the calls you missed! Only ¥9 for 3 months! Get ¥3 off if you reply “Y” to this SMS within the next 20
minutes!” Our corporate partner offered a ¥3 rebate to incentivize mobile users to respond. For those who
responded, the cost was charged to their phone bills. In our study, since the average subway commute is
30 minutes, we restricted the reply time to 20 minutes to ensure that most commuters would respond to
the SMS while still in transit.
3.2 Measuring Crowdedness
We measured crowdedness as the number of subway passengers per square meters. We did so in step-by-
step fashion. First, in our study, the subway consisted of articulated trains modeled after accordion-style
7
buses where the absence of doors between cars enables passengers to travel the length of the subway train
without restriction. Each subway train is composed of six articulated cars. We measured the volume of
mobile users in each subway train as the number of mobile users who are automatically connected to the
subway-tunnel cellular lines.2 Second, we calculated passenger volume. The mobile phone penetration
rate in major Chinese cities is very high. In the large city of our study, our corporate partner serves 70%
of the population. We thus derived passenger volume by dividing the mobile user volume by 0.7. Third,
we calculate crowdedness by dividing this passenger volume by the total area of each subway train. Each
subway car is 19 meters long and 2.6 meters wide, totaling 296.4m2 (6 cars x 19m length x 2.6m width).
3
3.3 Threats to Internal Validity
An ideal way to test the effects of crowdedness would be to conduct a randomized field experiment.
However, this is virtually impossible because it would require randomly assigning commuters to more
versus less crowded subway trains. Commuters are not accustomed to being assigned which subway train
they may ride. In other words, it is quite difficult to manipulate the level of crowdedness in a field setting.
Thus, our identification relies on field data. However, the internal validity of results based on field data
are often threatened by alternative explanations. We discuss each threat in turn, along with how we
address them to provide greater confidence in the results.
3.3.1 Threat Due to Self-Selection
The self-selection threat arises when people self-select onto more or less crowded trains depending on
their work schedule demands. For instance, during hours when there is likely to be more crowding, there
may be a greater proportion of working professionals than during other times, when more seniors,
children, and homemakers use the subway. If this is the case, the observed effects might well reflect
differences in behavior between people who travel during crowded rush hours versus other times, than the
effect of crowdedness per se. That is, purchases may be driven by variations in commuter type rather than
crowdedness. Thus, to reduce the concern of self-selection we deal with this threat in four specific ways.
First, crowdedness is more likely to occur during peak hours of travel. To isolate the effects of
crowdedness from the peak hours, we have field data across both peak and non-peak hours of
crowdedness. We do so by sending SMSs from morning to evening. For each day of our study, we
2 The wireless service provider can identify all mobile users’ phone numbers, including those of phones that are off (due to
telecommunication company chip-tracking technology). Thus, both the phones that are on and those that are off are recognized
and recorded. According to our corporate partner, over 99.9% of phones are on during the day.
3 Since crowdedness differs from city to city and culture to culture (e.g., cities with a high population density such as Hong Kong
are very different from those with a low population density such as Auckland), tolerance for crowdedness may also vary. Because
the Chinese are more tolerant of crowdedness given their large population, the effects of crowdedness in our field study may be
more conservative.
8
selected five different times to represent the five cycles people may experience during an average day.
Each time cycle represents a peak or non-peak hour of crowdedness and thereby potentially has a
different level of crowdedness. The first time cycle was from 07:30 - 08:30 and represents the morning
rush hour. The second was from 10:00 - 12:00, representing the lull after the morning rush hour traffic.
The third was from 14:00 - 16:00, the afternoon lull before the evening rush hour traffic. The fourth was
from 17:30 - 18:30, the evening rush hour. And the fifth was from 21:00 - 22:00, the after-dinner traffic.
Because time-cycles are nested within the trains, we control for subway train effects to obtain a more
finer-grained analysis. Thus, within the five time cycles, the corporate partner pushed SMSs to 14 subway
trains each day of the weekday and weekend parts of the study. These five time cycles nested in the 14
subway trains help control for the systematic differences between various types of subway commuters,
i.e., business commuters at rush hour versus non-business commuters at non-rush hours.
Second, crowding may differ across weekdays and weekends. On weekdays, work schedules might
force business travelers to self-select into crowded trains during rush hours, while non-business travelers
(e.g., retirees and homemakers) might self-select into non-crowded trains in lull hours. Also, on weekends,
leisure habits enable people to self-select into more or less crowded trains. Therefore, to isolate the effects
of crowdedness from different types of commuters across the weekday and weekend, we have field data
from two different days: a weekday (Wednesday) and a weekend (Saturday). These weekday and
weekend results help control for the systematic differences between weekday and weekend crowdedness
in our field data.
Third, although we do not have random treatment of crowdedness, in conjunction with our corporate
partner, we randomly selected subjects from among our targeted subway population in order to randomize
away user heterogeneity. Our targeted population had not previously subscribed to a missed call alert
plan, nor had it received a similar SMS from the corporate partner. This also helps to rule out potential
carryover effects of prior marketing campaigns. For example, if passengers have already received a
similar mobile promotion, it is possible that promotion repetition or fatigue drives the results. For each
user from this targeted subway population, we assigned a random number. We used SAS software’s
random number generator and ran the RANUNI function, which returns a random value from a uniform
distribution (Deng and Graz 2002). We then sorted the random numbers in sequence and extracted a
sample from the sequence. We integrated each of these steps into an algorithm in the wireless provider’s
IT system. This algorithm allowed us to compute and randomize users instantly in order to accurately
gauge the level of crowdedness while we sent the SMSs in real-time. Because the wireless service
provider maintains purchase records of all of its clients, we were able to know immediately whether a
given mobile user in the subway train had previously subscribed to the promoted service. Such dynamic
9
but instant computation and randomization are difficult to execute and require privileged support from the
corporate partner, thereby representing a unique feature of our field data.
Fourth, crowdedness may impact passengers differently depending on their mobile usage habits. For
example, users with more expensive phone bills might be more likely to purchase the missed call alert
package since they might connect with more people. Since our targeted purchasing behavior is for a
missed call-alert package, mobile usage can be reasonably expected to serve as good controls.
Specifically, we control for each mobile user’s wireless behavior, based on their monthly bills, call
minutes used, SMSs sent and received, and internet data usage.4 In the cell phone industry, ARPU, MOU,
SMS, and GPRS comprise key indicators of wireless usage behavior. ARPU (the average revenue per
user) is the revenue that one customer’s cellular device generated. MOU (individual monthly minutes of
usage) is how much voice time a user spent on her mobile. ARPU and MOU can help control for
customer heterogeneity, since business travelers are more likely to have a higher ARPU and MOU. In
addition, SMS (short message service) is the amount of monthly text messages sent and received. SMS
can also help tease out the effects of consumer age as younger generations are more prone to SMS usage.
GPRS (general packet radio service) is a measure of the individual monthly volume of data used with the
wireless service provider. GPRS is useful in terms of controlling for traveler habits in using the mobile
web and downloading mobile content. Table 1 provides summary statistics of these variables in natural
log to reduce the non-normality of the data. Specifically, panel A provides the statistics for the full sample
of these variables, panel B provides them for the combined part 1 and part 2 data (weekday and weekend
sample), and panel C provides them for the part 3 data (exogenous crowdedness sample). Appendix A
presents a histogram of ARPU for the full sample.
[Insert Table 1 Here]
Furthermore, we adopt two approaches to correct for selection bias and validate commuter
homogeneity. First, we analyze the effect of crowdedness with a subsample of the same-train-same-time
users. It is reasonable to expect that users in the same train at the same time during the weekday are
homogenous in their work schedules, i.e., similar types of passengers. Different types of passengers
would select different trains and different times to board. If so, analyzing the same-train-same time
mobile users helps establish commuter homogeneity. In addition, we used propensity score matching. We
mirrored randomization by matching the probability of mobile users entering the exogenous crowd versus
the non-exogenous crowd. In other words, using propensity score matching, our field data would be
equivalent to experimental randomization by transforming the field data into a quasi-experiment design
(Huang, Nijs, Hansen, and Anderson 2012; Rubin 2006). Thus, matching the treatment group with the
4 Government regulation prohibits the wireless provider from revealing customers’ private information so we are not able to
identify user demographics such as disposable income. However, income is not relevant in our context because the promoted
product costs about 50 cents per month.
10
control group in terms of their mobile usage behaviors enables us to better isolate the treatment effects of
crowdedness. This enables us to demonstrate that individuals who are virtually similar in all respects that
are measurable due to mobile usage behaviors vary their response to mobile targeting as a function of the
level of crowdedness.
3.3.2 Threat Due to Endogeneity
This threat occurs when the crowd is endogenous and when some unobserved variable may drive both
crowdedness and purchases. We deal with this threat by exploiting an exogenous shock to crowdedness.
Specifically, in part 3 of our field data, exogenous crowdedness resulted from a sudden variation in
crowdedness in the subway system. The local government temporarily controlled vehicular traffic for an
hour on a weekday. Because the traffic intervention was enforced to provide a high-security police escort
to government personnel, denizens were not forewarned. Thus, the temporary traffic intervention
exogenously changes the level of crowdedness within the subway trains, creating an exogenous spike in
crowdedness.5 Thus, our field data have consumer responses from both types of crowdedness; exogenous
and non-exogenous. If both types of crowdedness obtain directionally similar results, this would reveal
more evidence for the robustness of the crowdedness effect.
During the traffic intervention, the wireless provider sent SMSs to passengers on subway trains. We
promoted a different product category to help generalize the findings.6 The promoted service enabled
users to stream videos on their mobile devices. The SMS read “To watch the newest videos on your
mobile anytime, anywhere, for only 3¥ per month, reply KTV3 now and get 3¥ off of your next month’s
bill!” For mobile users who responded with the provided mobile short code, the cost was charged to their
phone bill. We measured crowdedness in the same manner and selected mobile users who had not
subscribed to this video service.
3.3.3 Threats Due to Other Factors
Several other factors may also threaten the validity of our results. We briefly explain them and how we
dealt with them. Differences in subway stations might present a threat. For example, some subway
stations have food stands and newspaper stands, while others do not. Also, passengers boarding at stations
5 By exogenous, we refer to the city’s denizens, rather than the corporate partner. In other words, the commuters did not know
ahead of time about the traffic intervention. An alternative way to derive exogenous crowdedness is to exploit train delays.
However, a single train delay would not be long enough to create a sufficient increase in crowdedness because every 3-5 minutes
a train enters a station. Also, a longer train delay is normally not feasible except for tunnel accidents or bomb threats (which
would create an opposite effect by forcing more people to use above-ground transit). Thus, while imperfect, this exogenous
crowdedness via the traffic intervention can also better control for self-section.
6 In the non-exogenously crowdedness sample, we promoted a missed call alert service. Yet, one may speculate that the presence
of many social others in the subway may prompt users to be more likely to subscribe to a missed call alert. Thus, in the sample of
exogenous crowdedness, we promoted a different service that has the same purchase price.
11
farther along a subway line experience less commute time than those who board earlier along the route. If
we sent mobile promotions to commuters located at different stations and traveling in different directions,
it is possible that station differences and observable commute length would affect the results. To reduce
these concerns, we pushed the SMS from only the fourth stop along the subway line. This helped to
ensure that enough passengers had boarded the train since the first station, but that they also still had a
sufficient distance to travel to stations further along the line. Also, the direction in which passengers
travel may also have an effect. If we sent mobile promotions to commuters traveling in different
directions, it is possible that commute direction would affect the results. Thus, we sent the SMS to trains
travelling in a single direction towards the city center.
Taken together, the above steps help reduce concerns of selection, endogeneity, and other threats, and
thereby help strengthen the case that we measure the effects of crowdedness.
3.4 Model
Our model estimates the purchase likelihood of each mobile user as . We
model the unobserved likelihood of making a mobile purchase as a logit function of crowdedness in
subway trains. Following Agarwal et al. (2011, p. 1063) and Guadagni and Little (2008), we assume an
i.i.d. extreme value distribution of the error term in the logit model:
=
( )
Crowd
iU = l+ l × crowdednessi +
l × subway_traini + l× Xi + i1
l, (1)
where Crowd
iU denotes the utility of a mobile purchase, l accounts for the unobserved fixed effects in
mobile users’ preferences for subway trains, and Xi is a vector of mobile user controls (natural logs of
ARPU, MOU, SMS, and GPRS). i is comprised of the idiosyncratic error terms. tests the effects of
crowdedness on mobile purchase probability after controlling for mobile user behaviors and subway train
fixed effects. We separate the effects of crowdedness from time cycles of the day by using subway train
fixed effects. Because subway train effects provide a finer-grained analysis of crowdedness than time
cycles (each time cycle contains several subway trains), we control for the effects of subway trains rather
than time cycles. Nevertheless, additional analyses show our results are robust to controlling for time
cycles. Furthermore, besides the train effects, crowdedness can also be related to weekday or weekend
effects (weekends and weekdays are expected to have different crowdedness patterns), so we further
identify the effects of crowdedness by controlling for weekday effects(m):
12
Crowd
iU = m+ m × crowdednessi + m× weekdayi + m
× subway_traini + m× Xi + i2 m
. (2)
Of key interest to us is the effect () of crowdedness on mobile purchase likelihood.
3.5 Results
3.5.1 Preliminary Evidence for the Effects of Crowdedness on Mobile Purchase Likelihood
We begin by reporting the response rate in our field data. Specifically, of the 10,690 mobile users who
received an SMS, 334 of them replied and purchased the promoted service, corresponding to a 3.22%
response rate in part 1 and 2 of our data. Of the 1,270 users who received an SMS during the traffic
intervention, 37 of them replied and purchased, equating to a 2.91% response rate in part 3 of our data.
These response rates seem low, but are fairly high compared to the 0.6% response rate for mobile
coupons in Asia (eMarketer 2012) and the 1.65% rate for the effectiveness of location-based mobile
coupons (Molitor et al. 2013). Appendix B presents the crowdedness and purchase rate per train.
A naïve way to test the effects of crowdedness on mobile purchase is to conduct analyses with
different samples. If the results for each sample of crowdedness in our data are statistically significant,
this provides initial evidence for the effect of crowdedness on mobile purchase. Table 2 presents this
preliminary evidence, after controlling for weekday versus weekend effects and peak hour versus non-
peak hour effects. In all models of Table 2, we first entered the four mobile usage covariates to control for
user heterogeneity in terms of observable mobile usage behaviors. We then entered the crowdedness of
each sample from our data in turn. As shown in Table 2, the effects of crowdedness are statistically
significant across the sample without the traffic intervention (non-exogenous weekday and weekend
crowdedness combined), the sample with the traffic intervention (exogenous crowdedness), and the full
sample consistently (all p < 0.05). As depicted in Figure 1, mobile purchase likelihood increases as a
function of crowdedness in the non-exogenous sample. As depicted in Figure 2, mobile purchase
likelihood increases as a function of crowdedness in the exogenous sample as well. This pattern gives
initial evidence for the effect of crowdedness.
[Insert Table 2, Figure 1, and Figure 2 Here]
3.5.2 Main Evidence for the Effects of Crowdedness on Mobile Purchase Likelihood
The preliminary evidence simply focuses on variations within the non-exogenously crowded commuters
alone, within the exogenously crowded commuters alone, or within the full sample of commuters. Yet,
the results may be confounded by sample selection and endogeneity threats. As commuters may self-
select into different trains, the crowd may be endogenously induced, and some unobserved variable may
drive both crowdedness and purchases and hence drive the results in all the samples. Thus, we provide
13
evidence in dealing with these threats by exploiting an exogenous shock to crowdedness. Specifically, we
pool the exogenous traffic intervention sample and the non-exogenous sample, and then add the
interaction between the traffic intervention and crowdedness. This interaction identifies the difference in
behavior between commuters in the exogenously crowded trains and those in the non-exogenously
crowded trains. If this interaction is significant, that would be a stronger test of the impact of crowdedness
because this impact is driven by the exogenous crowdedness free from endogeneity bias. Thus we
developed the following model to identify that variations in mobile purchases stem from variations in
Ln(GPRS) 8.8505 2.2358 6.5089 7.7497 8.1521 8.4282 9.0002 9.4204 9.7947 10.1545 10.3388 10.4607 10.8110 Note: ARPU, MOU, SMS, and GPRS comprise key indicators of wireless usage behavior. ARPU (the average revenue per user) is the revenue that one customer’s cellular device generated. MOU
(individual monthly minutes of usage) is how much voice time a user spent on her mobile. SMS (short message service) is the amount of monthly text messages sent and received. GPRS (general
packet radio service) is a measure of the individual monthly volume of data used with the wireless service provider.
23
Table 2: Preliminary Evidence for Effect of Crowdedness on Mobile Purchase
Parameter
Sample Without
Traffic
Intervention
Sample Without
Traffic
Intervention
Sample With
Traffic
Intervention
Sample With
Traffic
Intervention Full Sample Full Sample
Crowdedness 0.114**
(0.042)
0.601**
(0.285)
0.125**
(0.041)
Ln(ARPU) 0.305**
(0.128) 0.305**
(0.128) 0.239
(0.339) 0.236
(0.337) 0.301**
(0.118)
0.299**
(0.118)
Ln(MOU) -0.043
(0.070) -0.044
(0.069) 0.020
(0.220) 0.018
(0.217) -0.043
(0.065)
-0.044
(0.065)
Ln(SMS) 0.003
(0.073) 0.006
(0.073) 0.146
(0.250) 0.124
(0.253) 0.014
(0.069)
0.015
(0.069)
Ln(GPRS) -0.001
(0.025) -0.001
(0.024) 0.004
(0.086) 0.008
(0.085) -0.001
(0.024)
0.000
(0.023)
Day Effects
(Weekend
Dummy)
Yes Yes Yes Yes Yes Yes
Time Effects
(Peak Hour
Dummy)
Yes Yes Yes Yes Yes Yes
Observations 10,690 10,690 1,270 1,270 11,960 11,960 Note: **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS = number of texts sent and received per user, GPRS
= data usage with the wireless provider.
Table 3: Evidence for Effect of Crowdedness on Mobile Purchase
Parameter Model 1 Model 2 Model 3 Model 4
Crowdedness
X
Traffic Intervention
0.492**
(0.187)
Crowdedness 0.126**
(0.041)
0.114**
(0.042)
Traffic Intervention -0.120
(0.117)
-0.142
(0.177)
-1.887
(1.057)
Ln(ARPU) 0.301**
(0.118)
0.308**
(0.119)
0.308**
(0.119)
0.306**
(0.119)
Ln(MOU) -0.043
(0.065)
-0.043
(0.065)
-0.044
(0.065)
-0.044
(0.065)
Ln(SMS) 0.014
(0.069)
0.014
(0.069)
0.015
(0.069)
0.013
(0.069)
Ln(GPRS) -0.001
(0.024)
-0.001
(0.023)
-0.001
(0.023)
-0.001
(0.023)
Day Effects (Weekend
Dummy) Yes Yes Yes Yes
Time Effects (Peak Hour
Dummy) Yes Yes Yes Yes
Observations 11,960 11,960 11,960 11,960 Note: **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS =
number of texts sent and received per user, GPRS = data usage with the wireless provider.
24
Table 4: Robustness Checks with Same-Train-Same-Time Homogenous Sample
Parameter Model 1 Model 2 Model 3 Model 4
Crowdedness
X
Traffic Intervention
0.430***
(0.103)
Crowdedness 0.124**
(0.041)
0.116**
(0.042)
Traffic Intervention -0.075
(0.081)
-0.073
(0.081)
-0.042
(0.083)
Ln(ARPU) 0.301**
(0.118)
0.309**
(0.118)
0.307**
(0.118)
0.305**
(0.118)
Ln(MOU) -0.043
(0.065)
-0.040
(0.065)
-0.041
(0.065)
-0.040
(0.065)
Ln(SMS) 0.014
(0.069)
0.000
(0.070)
0.002
(0.070)
-0.001
(0.070)
Ln(GPRS) -0.001
(0.024)
0.002
(0.024)
0.003
(0.024)
0.003
(0.024)
Observations 1,550 1,550 1,550 1,550 Note: **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS = number
of texts sent and received per user, GPRS = data usage with the wireless provider.
Table 5: Robustness Checks with Propensity Score Matching-based Homogenous Sample
Parameter Model 1 Model 2 Model 3 Model 4
Crowdedness
X
Traffic Intervention
0.318***
(0.075)
Crowdedness 0.131**
(0.052)
0.125**
(0.047)
Traffic Intervention -0.098
(0.073)
-0.082
(0.069)
-0.095
(0.076)
Ln(ARPU) 0.133
(0.156)
0.134
(0.157)
0.130
(0.155)
0.133
(0.152)
Ln(MOU) -0.105
(0.083)
-0.107
(0.085)
-0.101
(0.083)
-0.103
(0.087)
Ln(SMS) 0.004
(0.086)
0.006
(0.082)
-0.003
(0.091)
-0.004
(0.093)
Ln(GPRS) -0.012
(0.030)
-0.015
(0.032)
-0.011
(0.031)
-0.016
(0.033)
Observations 782 782 782 782 Note: ***p < 0.01; **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS =
number of texts sent and received per user, GPRS = data usage with the wireless provider.
25
Table 6: Residual Approach to Robustness Checks
Panel A: Crowdedness as Dependent Variable
Parameter Model 1
Ln(ARPU) 0.026
(0.027) Ln(MOU) 0.018
(0.015) Ln(SMS) -0.031**
(0.016) Ln(GPRS) -0.003
(0.005)
Observations 2,886
Panel B: Mobile Purchase as Dependent Variable
Parameter Model 1 Model 2 Model 3 Model 4
Ln(ARPU) 0.301**
(0.118)
0.310**
(0.118)
0.309**
(0.118)
0.307**
(0.128)
Ln(MOU) -0.043
(0.065)
-0.039
(0.065)
-0.038
(0.065) -0.039
(0.069)
Ln(SMS) 0.014
(0.069)
-0.002
(0.070)
-0.004
(0.071) -0.011
(0.077)
Ln(GPRS) -0.001
(0.024)
0.002
(0.024)
0.002
(0.024) 0.002
(0.025)
Traffic Intervention -0.073
(0.081)
-0.056
(0.085) -0.067
(0.131)
Residual_Crowdedness 0.124**
(0.041)
0.128**
(0.042) 0.126**
(0.042)
Residual_Crowdedness X Traffic Intervention 0.416***
(0.132) 0.412**
(0.142)
Residual_Crowdedness X Weekend -0.045
(0.107)
Residual_Crowdedness X Peak Hour -0.001
(0.015)
Day Effects (Weekend Dummy) Yes Yes Yes Yes
Time Effects (Peak Hour Dummy) Yes Yes Yes Yes
Observations 11,960 11,960 11,960 11,960 Note: **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS = number of texts sent and received per user,
GPRS = data usage with the wireless provider.
26
Table 7: Falsification Tests
Subsample with Low Crowdedness (under 2
passengers/m2), Lower Boundary
Full Sample with Other
Interactions Parameter Model 1 Model 2 Model 3
Ln(ARPU) 0.354
(0.267)
0.352
(0.265)
0.299
(0.321) Ln(MOU) -0.235
(0.142)
-0.231
(0.142)
-0.227
(0.179) Ln(SMS) 0.083
(0.146)
0.074
(0.148)
0.107
(0.193) Ln(GPRS) -0.028
(0.052)
-0.026
(0.053)
-0.003
(0.068)
Crowdedness -0.084
(0.270)
0.121**
(0.316)
Crowdedness x Ln(ARPU) 0.002
(0.088)
Crowdedness x Ln(MOU) 0.054
(0.049)
Crowdedness x Ln(SMS) -0.027
(0.054)
Crowdedness x Ln(GPRS) 0.001
(0.018)
Day Effects (Weekend Dummy) Yes Yes Yes
Time Effects (Peak Hour Dummy) Yes Yes Yes
Observations 2,886 2,886 11,960
Note: **p < 0.05; *p < 0.10; ARPU = average revenue per user, MOU = minutes of usage, SMS = number of texts sent and received per user, GPRS = data